LTX-2.5 in ComfyUI: Files, Settings and the $10M Licence Rule

LTX-2.5 runs in ComfyUI through native templates, and the smallest set of files for text-to-video or image-to-video with sound is 39.71 GB. Lightricks’s own system requirements ask for an NVIDIA GPU with at least 32GB of VRAM. The weights are gated on Hugging Face, and the licence is free only below $10 million in annual revenue.

This guide gives the exact files and folders for each template, their settings, what is stated about memory, how sound works, the errors people report, which IC-LoRAs are worth downloading, and what the licence says. For a comparison with Wan and HunyuanVideo, see our ComfyUI video generation guide. BFWR has not run LTX-2.5. Everything here comes from Lightricks’s documentation and repository, ComfyUI’s documentation, source and template files, the published file listings, and public GitHub issues.

Quick answer

• Smallest working download: 39.71 GB for text-to-video or image-to-video, 38.71 GB for first-and-last-frame. The prompt enhancer adds 5.20 GB. The higher-quality DFR template needs 51.19 GB, or 61.47 GB with its enhancer.
• Memory: Lightricks’s documentation says 32GB+ VRAM, while its marketing page says 16GB. ComfyUI publishes no figure.
• Settings: the distilled templates use 8 steps, then 3, at CFG 1, on a fixed sigma schedule. Keep it.
• Licence: entities with annual revenue of $10 million or more need a paid licence for anything beyond narrow non-commercial use.

Get access before you download

ComfyUI’s tutorial is direct about this: “Open the repository pages, accept the model license, and wait for your access request to be approved before downloading the models.” You need that for two repositories, Lightricks/LTX-2.5 and, for the DFR template only, Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler. The optional enhancer sits in Comfy-Org/gemma-4, which is not gated.

Update ComfyUI first. The template index sets a minimum of version 0.32.0 (released 11 August 2026) for the three basic templates and 0.37.2 for DFR. On 13 November 2026 the latest release is v0.45.0, from 10 November. New to it? See our ComfyUI install guide.

In ComfyUI, open Templates, search LTX-2.5 and press Download all. To download by hand, use the Hugging Face CLI, which keeps the repository’s folder layout; those folders match ComfyUI’s. This is Lightricks’s README command with the file names changed to the ComfyUI builds:

hf auth login
hf download Lightricks/LTX-2.5 \
    diffusion_models/ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors \
    text_encoders/gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors \
    vae/ltx-2.5-video-vae-bf16.safetensors \
    vae/ltx-2.5-audio-vae-bf16.safetensors \
    latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
    --local-dir ComfyUI/models

A 401 or 403 means the terms are not accepted or the token is wrong. Lightricks says to log in with a Read token; fine-grained tokens need the read-gated-repos scope.

Exact files and folders

Sizes are from the Hugging Face file listings on 9 November 2026, where GB means a billion bytes. The listing is readable even though the weights are gated.

These totals are the sum of the listed files. ComfyUI’s template index shows larger figures for the first three (53.15 GB for T2V, 56.91 GB for I2V and FLF2V), which do not match the files they load.

Both projects call the enhancer optional, but each template still contains its CLIPLoader, and ComfyUI’s validator (execution.py) checks every linked node before a run. So a missing enhancer file looks likely to stop the queue even with the toggle off. If it does, download it or delete that branch.

Do not swap in a generic Gemma file: Lightricks’s README says “Google’s stock Gemma 4 release is not a substitute”, because loading checks the encoder’s version. The comfy-int8-convrot builds are for ComfyUI only. The full BF16 stack in the README comes to roughly 66 GiB.

Settings the templates ship with

These come from the template JSON files. No template needs a custom node pack in current ComfyUI: every node, including DFR’s keyframe nodes in comfy_extras/nodes_lt_keyframes.py, is in core. The index still tags DFR with requiresCustomNodes, which appears to be out of date.

Stage 1 of T2V and I2V uses ManualSigmas of 1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0, and stage 2 uses 0.85, 0.7250, 0.4219, 0.0. ComfyUI’s DFR page explains why to leave them: “The checkpoint is distilled, so the step counts and the schedules are part of the recipe”. Video decoding uses VAEDecodeTiled to hold down VRAM. Image-to-video uses LTXVImgToVideoInplace at strength 0.7, then 1; FLF2V uses LTXVAddGuide at frames 0 and -1, both at 0.7.

Two rules for size and length: width and height move in steps of 32, and frame count must be 1 plus a multiple of 8. T2V and I2V compute frames as duration times fps plus 1, so 24 fps with whole seconds is always valid, while 25 fps for 5 seconds gives 126, which is not.

Lightricks’s guide lists the T2V default as 768×512 and 97 frames with Prompt Enhance on. The current template JSON uses the values above with the enhancer saved off, as ComfyUI’s tutorial says. Turned on, it “adds about 1-2 minutes of generation time”.

How much VRAM it needs

The official statements disagree. Lightricks’s system requirements page says “NVIDIA GPU with a minimum 32GB+ VRAM – more is better”, with 32GB of system RAM, 100GB free disk, CUDA 12.7 or later and Python 3.12 or later, and it recommends an A100 (80GB) or H100. The LTX-2.5 product page, in a table comparing it with rivals, gives a 16GB minimum. ComfyUI’s template index leaves vram empty, and nobody publishes memory by resolution.

The file arithmetic, which is not a measurement: the int8 transformer and int8 text encoder total 36.87 GB, so both cannot sit on a 32GB card at once, and ComfyUI loads them in stages. A log in issue #15606 from an RTX 5090 shows 14,612 MB staged for the encoder and 20,484 MB for the transformer. In #15600 the owner of an RTX 5070 Ti, a 16GB card, reports good results at 1280×720. That is a user report, not a guarantee.

Short of memory, Lightricks suggests the int8 builds, its lighter ltx-2.5-video-vae-conv-bf16.safetensors (1.45 GB; the templates default to the higher-quality decoder), and testing at low resolution with 41 to 81 frames. Our 32GB VRAM tier page shows what else fits that class of card.

Sound

LTX-2.5 generates audio and video together in stage 1 and carries the audio through later stages. There is no separate audio model, but the audio VAE, ltx-2.5-audio-vae-bf16.safetensors, is a separate file every template loads. For speech, describe it in the prompt and put spoken lines in quotation marks.

Two more audio workflows sit outside the template library. Lightricks’s LTX-2.5_A2V_Two_Stage_Distilled.json, in the ComfyUI-LTXVideo repository, holds your own clip fixed and generates video to match. Lip-sync to custom audio is the weak spot: in #15643 users report mouths out of sync on graphs that worked on 2.3, and the stale bot closed it without a fix.

Common failures and what to do

• Prompt enhancer returns nothing or nonsense. TextGenerateLTX2Prompt can return an empty string on the text-only path (#15596), output the prompt twice (#15649, 4 of 24 image-fed seeds in one reproduction) or drift off the subject (#15600). All open. Leave it off, or read the preview node’s text before a long render.
• Windows fatal exception: access violation after sampling. On a 32GB card, the process died while building the diffusion VAE decoder with the transformer still in memory (#15606, open). A model-unload node between sampler and decode got the reporter through; smaller renders did not trigger it.
• RuntimeError: mat1 and mat2 shapes cannot be multiplied. Reported with ltx-2.5-22b-distilled-transformer-nvfp4.safetensors through the stock UNETLoader (#15511, open). Use the int8 file.
• ValueError: Unexpected numpy array shape in LTXVPreprocess. The input image has an alpha channel (#16679, open). Save it as RGB.
• Own graph, empty audio VAE list. Use a plain VAELoader and CLIPLoader type ltxv, as the templates do. LTXVAudioVAELoader reads models/checkpoints/, not models/vae/ (#15595).
• cannot import name 'pad' from 'kornia.geometry.transform.pyramid'. ComfyUI-LTXVideo fails to load with newer kornia (#542, open). Commenters report kornia==0.8.2 works. It affects IC-LoRA workflows, not the native templates.
• GGUF size mismatch. Community builds through UnetLoaderGGUF have failed this way (#15540).
• Mac and AMD. Lightricks lists NVIDIA only. On Apple Silicon, #15818 says --use-split-cross-attention cured all-black video on an M1 Max; #15804 records 19 of 26 black runs on an M5 Max. On an RX 7600 XT, LTX-2 #325 describes driver timeouts and full freezes with the stock template at 864×480.

ComfyUI binds to 127.0.0.1 by default. Leave it there unless you mean to expose it; --listen with no address opens all interfaces.

Which IC-LoRAs are worth knowing

IC-LoRAs take a reference video or control signal as well as a prompt, and go in models/loras/. Their workflows need the ComfyUI-LTXVideo pack. Every repository below is gated, so the model cards return 401 without an account; this table is from Lightricks’s documentation and the file listings. Read each card after accepting its terms, because it sets the loader, prompt wording and strength.

After the upscaler, Refine Details is the one most people should try. Lightricks’s warning on the restoration pair applies to all of them: “They synthesize plausible color and detail; they do not recover the source pixel-for-pixel.” Refine Details may also remove original film grain. Ingredients has a trap: Lightricks says its workflow currently loads the LTX-2.3 checkpoint while the 2.5 card names a separate one, and an open report (#586) shows a four-view sheet staying as four figures in the output.

The README says “a LoRA only works with the model it was trained on”, while Lightricks’s LoRA guide says most 2.3 adapters run on 2.5, with exceptions. Use the exact file a workflow names.

The $10M licence rule

The code in Lightricks/LTX-2 is Apache 2.0; the weights are not. They are under the LTX-2.x Community License, dated 11 August 2026 and amended on 29 October 2026 to separate the Apache-licensed code from the weights and to extend the disclosure rule to whatever software runs the model. This is a reading of the published text, not legal advice.

• The threshold. Section 2.1 grants worldwide use for any purpose, except that “Entities with annual revenues of at least $10,000,000” need a paid licence. The product page says “Free under $10M ARR”, but the licence says annual revenues, so go by that. Below the line, commercial use is allowed, including paid hosting under section 3.
• Group revenue counts: “all subsidiaries, affiliates, and companies under common Control shall be considered collectively.”
• The narrow exception. Above the line, a company may still use it for “testing, evaluation, or non-commercial research and development in a non-production or development environment”, but not “in direct interactions with or that has impact on end users”.
• Your outputs. “Licensor claims no rights in the Output you generate”, but outputs stay subject to the use restrictions.
• LoRAs inherit the licence. Section 3.5 names “fine-tuned weights, LoRA adapters, or similar adaptations” as Derivatives, and section 3.2 adds “models trained on Outputs”. They “must be distributed exclusively under the terms of this Agreement”, and you may not hand one to a Commercial Entity unless it holds a paid licence.
• Do not strip disclosures. Section 6 forbids removing “safety or security measures, disclosures, metadata, watermarking”, and Attachment A bars publishing generated content “without expressly and intelligibly disclaiming that the information and/or content is machine generated”.
• Do not compete with Lightricks. Attachment A item 20 bars use in “any product, service, or application that directly competes with Licensor’s commercial products or services” without a separate licence. If you are building a video tool, read it twice.

There is no territory exclusion. The text does not say whether a small studio working for a large client falls under the client’s revenue, so ask a lawyer if that is you. Our Flux LoRA licence guide shows how another model treats the same questions.

Is it maintained?

Yes. In the 90 days to 9 November 2026, Lightricks/LTX-2 had 15 commits (last on 4 November) and Lightricks/ComfyUI-LTXVideo had 24 (last on 20 October, release v1.1.0). Neither is archived.

FAQ

Do I need ComfyUI-LTXVideo for LTX-2.5?

Not for the four native templates, whose nodes are in ComfyUI core. You need it for IC-LoRA and the other example workflows in that repository.

Can LTX-2.5 run on 16GB of VRAM?

Lightricks’s product page says 16GB and its system requirements say 32GB or more. Users report both success and crashes on 16GB cards, so start at low resolution.

Can I sell videos I make with LTX-2.5?

Below $10 million in annual revenue, the licence allows commercial use subject to its restrictions, and Lightricks claims no rights in your outputs. At or above that level you need a paid licence.

Related guides and tools

• Local AI video generation in ComfyUI: Wan 2.2, HunyuanVideo 1.5 and LTX-2.5
• ComfyUI install: Desktop, Portable or manual
• ComfyUI templates: which run locally and which need the cloud
• Best GPU for AI in 2026
• How to increase VRAM
• Flux LoRA licence guide
• ComfyUI models and VRAM
• Best local models for 32GB VRAM
• Model licence checker

Sources

• Lightricks/LTX-2 on GitHub (README) and its LTX-2.x Community License
• Lightricks/LTX-2.5 on Hugging Face (gated; file listing and API metadata only) and Comfy-Org/gemma-4
• ComfyUI: LTX-2.5 tutorial and LTX-2.5 DFR
• Comfy-Org/workflow_templates (index and the four template JSON files)
• ComfyUI source (comfy_extras/nodes_lt.py, nodes_lt_keyframes.py, nodes_lt_audio.py, nodes_resolution.py, execution.py, comfy/cli_args.py)
• Lightricks documentation: system requirements, ComfyUI setup, text-to-video, audio-to-video, IC-LoRA adapters, LoRAs, Refine and Restore and the prompting guide
• LTX-2.5 product page (16GB and $10M ARR claims)
• GitHub issues as linked above, in Comfy-Org/ComfyUI, Lightricks/ComfyUI-LTXVideo and Lightricks/LTX-2

Updated 9 November 2026. File sizes, template settings, licence text, issue status and repository activity were read on that date. Re-checked 13 November 2026: licence text and file sizes unchanged; the DFR template now requires ComfyUI 0.37.2.